Research On MCMC Based Object Tracking Method

Object tracking is an important research division in the field of computer vision that it can be used in various application such as artificial intelligence, human- machine interaction, automatic system and etc. Object tracking methods server as the fundamental role in the division of computer vision and make an outstanding significance in higher lever research area such as human action recognition and comprehension. Object tracking relevant algorithm could divided into two different types: one kind is so-called discriminant tracking methods and another kind is subjected to random sampling based methods. Discriminant methods perform badly while handling with tracking scenarios containing abrupt scale changes and position changes. Aimed at such targets, random sampling methods based on Bayesian tracking scheme have achieved a lot recently. Inadequate researches focused on such area during past years and it’s particularly significant as high level tracking goals approximating to real scenarios are raised to adapt to our rapidly changing and developing life and technology better.This paper contributes to push research on Markov C hain Monte Carlo random sequential sampling based on Bayesian tracking scheme, especially to scenarios with abrupt position changes and scale changes and proposed a new tracking algorithm named Multi-Scale MCMC-Based Tracking Method. Based on traditional MCMC tracking method, weighted proposal method and MeanShift optimization method are added to tracking frame and as a result we improve the accuracy rate efficiently.Another contribution is combining compressive sensing with Markov Chain Monte Carlo tracking scheme to achieve better accuracy and results. On the basis of experiments, we observed there exist multi-peak Gussian distribution around the ground truth. The method we proposed can handle with local optimal solution that the traditional search-detect method may fall in and improved precision finally.